> Seems strange that this requires java.

A lot of deployed deep learning is in Python so I doubt this will matter. Looking at the docs, it looks like the main speech to text is python heavy, it leverages elasticsearch and calls libraries that are bindings to code written in C++. As none of that overhead will ultimately matter for anything approaching the fidelity you get with cloud services, cost benefit trade off favors Java in this scenario.

> lower profile computers

When it comes to deep learning, this just isn't the direction things are headed. If you want the best results with low latency, your home server is going to have to be fairly beefy with a decent GPU and lots of memory.